Satellite configurations for Besicovitch covering lemma in vector spaces #
THIS FILE IS SYNCHRONIZED WITH MATHLIB4. Any changes to this file require a corresponding PR to mathlib4.
The Besicovitch covering theorem ensures that, in a nice metric space, there exists a number N
such that, from any family of balls with bounded radii, one can extract N families, each made of
disjoint balls, covering together all the centers of the initial family.
A key tool in the proof of this theorem is the notion of a satellite configuration, i.e., a family
of N + 1 balls, where the first N balls all intersect the last one, but none of them contains
the center of another one and their radii are controlled. This is a technical notion, but it shows
up naturally in the proof of the Besicovitch theorem (which goes through a greedy algorithm): to
ensure that in the end one needs at most N families of balls, the crucial property of the
underlying metric space is that there should be no satellite configuration of N + 1 points.
This file is devoted to the study of this property in vector spaces: we prove the main result
of Füredi and Loeb, On the best constant for the Besicovitch covering theorem,
which shows that the optimal such N in a vector space coincides with the maximal number
of points one can put inside the unit ball of radius 2 under the condition that their distances
are bounded below by 1.
In particular, this number is bounded by 5 ^ dim by a straightforward measure argument.
Main definitions and results #
multiplicity Eis the maximal number of points one can put inside the unit ball of radius2in the vector spaceE, under the condition that their distances are bounded below by1.multiplicity_le Eshows thatmultiplicity E ≤ 5 ^ (dim E).good_τ Eis a constant> 1, but close enough to1that satellite configurations with this parameterτare not worst than forτ = 1.is_empty_satellite_config_multiplicityis the main theorem, saying that there are no satellite configurations of(multiplicity E) + 1points, for the parametergood_τ E.
Rescaling a satellite configuration in a vector space, to put the basepoint at 0 and the base
radius at 1.
Disjoint balls of radius close to 1 in the radius 2 ball. #
The maximum cardinality of a 1-separated set in the ball of radius 2. This is also the
optimal number of families in the Besicovitch covering theorem.
Any 1-separated set in the ball of radius 2 has cardinality at most 5 ^ dim. This is
useful to show that the supremum in the definition of besicovitch.multiplicity E is
well behaved.
If δ is small enough, a (1-δ)-separated set in the ball of radius 2 also has cardinality
at most multiplicity E.
A small positive number such that any 1 - δ-separated set in the ball of radius 2 has
cardinality at most besicovitch.multiplicity E.
Equations
- besicovitch.good_δ E = _.some
A number τ > 1, but chosen close enough to 1 so that the construction in the Besicovitch
covering theorem using this parameter τ will give the smallest possible number of covering
families.
Equations
- besicovitch.good_τ E = 1 + besicovitch.good_δ E / 4
Relating satellite configurations to separated points in the ball of radius 2. #
We prove that the number of points in a satellite configuration is bounded by the maximal number
of 1-separated points in the ball of radius 2. For this, start from a satellite congifuration
c. Without loss of generality, one can assume that the last ball is centered at 0 and of
radius 1. Define c' i = c i if ‖c i‖ ≤ 2, and c' i = (2/‖c i‖) • c i if ‖c i‖ > 2.
It turns out that these points are 1 - δ-separated, where δ is arbitrarily small if τ is
close enough to 1. The number of such configurations is bounded by multiplicity E if δ is
suitably small.
To check that the points c' i are 1 - δ-separated, one treats separately the cases where
both ‖c i‖ and ‖c j‖ are ≤ 2, where one of them is ≤ 2 and the other one is > 2, and
where both of them are > 2.
In a normed vector space E, there can be no satellite configuration with multiplicity E + 1
points and the parameter good_τ E. This will ensure that in the inductive construction to get
the Besicovitch covering families, there will never be more than multiplicity E nonempty
families.